Higgsfield's CEO: $4M a Month Burned on AI Models

Why Higgsfield's $4M Monthly Model Bill Is the New Cost of Knowledge Work

Alex Mashrabov, CEO of AI video company Higgsfield, sat down with Harry Stebbings to explain how a company that burned through most of its first $16M round finding fit reached $1B in annualized revenue in 18 months. The numbers that matter for anyone deploying AI agents are not the revenue ones. They are the cost lines: what it costs to run a company where every employee works through models.

The internal bill: "At Higgsfield, a team member spends an average of more than $10,000 a month on various models... Internal usage of models per month is over 4 million." With roughly 400 people, model spend now sits next to payroll as a core input. One creative spent over $30,000 in a single week live-coding a replacement for the company's asset workflow. It was not production-ready, but Mashrabov counts it as a positive experiment.

Where it goes next: Mashrabov expects the best engineers and creatives ("10X creatives") to reach $50,000-$100,000 a month in model spend, while roles like legal and finance plateau far lower. He also expects those people to ask for matching pay. The spend is becoming a per-role productivity budget, not an IT line item.

Tokenomics and routing: "The margin in open-weight models exceeds 80%... for closed-source models, it's probably between 20% and 30%." Higgsfield's agentic workflows let the company choose the model in more than 40% of cases, so routing each job to the cheapest model that does it well becomes a core feature. His reasoning: "You don't necessarily need a PhD level of intelligence to make viral videos on social media."

Agents don't remove the human teams (yet): He assumed legal and customer service would be largely replaced. Instead legal has 10+ people and customer success 40+, all using AI heavily. He estimates over 60% of front-line support requests can be handled by AI, but B2B is harder, and new products ship weekly, so the agents' context and rules go stale fast.

Benchmarks vs. real work: Video benchmarks test text-to-video, but real Higgsfield scenes use prompts averaging over 3,000 words and at least 10 image references. His conclusion is that benchmarks "do not represent that" workflow, and that people optimizing for them are not measuring production usefulness.

Moats in the agent era: He sees only two: delivering the outcome (for Higgsfield, ads that sell) and network effects, which AI does not replace. Community open-source projects on the platform grew from about 10 to over 10,000 in eight weeks.

5 Takeaways from Mashrabov on Running a Company on AI Spend

• Model spend is the new headcount line - $10k+ per person per month at Higgsfield, with a path to $50-100k for top performers.
• Route by task, not by brand - open-weight and post-trained models earn 80%+ margins versus 20-30% on closed models, so routing is a margin lever.
• Usage-based expansion beats seat pricing - one customer went from $99 a month to a $6M annual deal in six months, and month-12 net revenue retention exceeds 300%, though about 30% of users drop in month one.
• Humans stay in the loop for now - legal and support teams grew in headcount while adopting AI, because agents need constantly updated context and rules.
• Coding tools switch fast - engineers moved from Claude to Codex in mid-June, and creatives followed, so tool loyalty is cyclical.

What This Means for AI-Powered Organizations

The Higgsfield numbers are a preview of agent economics: the marginal cost of knowledge work moves from salaries to tokens, and the winners manage that spend like a portfolio, routing cheap models to routine work and reserving frontier models for hard problems. Treat the claims as one founder's self-reported figures; the revenue method is run-rate (last four weeks x 13), not audited annual sales. Still, the pattern of per-employee model budgets and model routing is what any organization running AI agents should plan for.

Watch the discussion of open-weight margins and model routing at 30:50